Structural Image Analysis in Epilepsy

Magnetic resonance imaging (MRI) has revolutionized the management and understanding of epilepsy. Routine inspection of high-resolution MRI allows identification of lesions in many patients with epilepsy, particularly those with refractory epilepsy. There remain ∼20% of patients with refractory partial seizures in whom even high-resolution MRI fails to demonstrate abnormality 1. The detection of more subtle abnormalities using MRI is therefore a clinical imperative and has led to new MRI acquisition, image processing, analysis, and objective quantification methods. We describe novel techniques affording improved ability and reliability of detection of abnormalities in epilepsy, some of which have already demonstrated their potential. When the brain is viewed in serial sections, its three-dimensional (3D) gyral structure is not easily appreciable: data can be misinterpreted. Evaluation of cortical thickness and the gray–white matter interface is often impaired by sectional obliquity in relation to gyral folding. These limitations often prevent adequate evaluation of gyral structure in localisation-related epilepsy. Previous attempts have been made to present data in nonorthogonal sections. Multiplanar reformatting permitted by volumetric data may reveal subtle abnormalities 2. Three-dimensional surface rendering facilitates examination of realistic surface gyral patterns, may enable identification of gyral abnormality 3, a biologic feature of many epileptogenic pathologies, and may aid surgical management 4. Curvilinear multiplanar reformatting (CMPR) 5 was developed to overcome some limitations of 3D reconstruction and rectilinear planar reformatting. CMPR allows the creation and viewing of curved slices along the hemispheric convexities. In CMPR, the distance of the serial curved slices from the curved brain surface is maintained constant throughout the cortical ribbon. The images obtained are effectively concentric, reducing artefactual cortical thickening secondary to obliquity of section plane. "Surface" topography is better maintained, with improved visualization of gyral and sulcal dimensions and spatial arrangement. The images allow comparison of morphologic details between adjacent and homologous gyri. CMPR (Fig. 1) has identified subtle dysplastic lesions in patients whose conventional MRIs were considered normal 5. Subtle, small areas of cortical thickening and gray–white matter interface blurring may be revealed. Improved morphologic characterization of various structural lesions also is possible with better delineation of lesion extent, anatomic localization, and anatomic relation of lesions to eloquent cortex. CMPR images also are amenable to co-registration with various functional data, including positron emission tomography (PET), functional MRI (fMRI), and stimulation studies, aiding clinical interpretation and presurgical planning. Subtle focal cortical dysplasia. A: Curvilinear multiplanar reformatting (CMPR) image obtained at the brain–cerebrospinal fluid interface (0 mm) demonstrates sulcal asymmetry of the right first frontal gyrus (arrow). This finding becomes more apparent at a deeper level (B, 4 mm from surface). C: The slice obtained at 12 mm from the surface shows the right first frontal gyrus split in half by a deep sulcus. The cortex of that gyrus is slightly thicker when compared with adjacent gyri, and the interface between gray and white matter is blurred. D: A slice at 16 mm from the surface shows the bottom of the aforementioned gyrus. The cortex is thicker at this level and the blurring of gray–white matter transition is more apparent. Observe that the anatomic display provided by CMPR allows accurate localization of the lesion and assessment of the anatomic relation between the lesion and the precentral sulcus (arrowheads). Segmentation is the identification of "natural" structures in an image. It can be done with various degrees of automation and objectivity. The delineation of the hippocampi for volumetric measurements has been the most important application of image analysis in epilepsy research. The hippocampus is a small, anisotropic, elongated, and poorly differentiated formation from neighbouring structures on MRI. These factors combine to render automation of its segmentation difficult. At the other end of the spectrum, the brain as a whole is of interest for a number of applications, including morphologic studies of the cortex, volumetry for monitoring disease progression and correction of hippocampal volumes for age and gender. Various methods are currently available to segment the whole brain, or its constituent gray and white matter (GM and WM) automatically, based on multiecho or single-echo volumetric data with a degree of precision of the order of 1%6. One method of segmentation of GM, WM, and cerebrospinal fluid (CSF) in a fully automatic fashion allows objective and precise calculation of the intracranial volume 7(Fig. 2). Automatic compartmentalization of the brain into its main natural subdivisions remains a long-term aim. Currently, a promising approach to this problem is the mapping of digital atlases onto an individual subject's image data set using (nonrigid) registration. This is complicated by the brain's morphologic variability, including the presence of topologic differences between individuals and the atlas (representative or symbolic image) that are difficult to reconcile. Nonetheless, substantial progress has been made, and applications are emerging. Illustration of serial magnetic resonance imaging registration and fully automatic segmentation of the brain. The two scans were acquired 9 months apart. The segmentation result is shown as a white outline. Image registration, the mapping of homologous regions between imaging studies, is a prerequisite for many MRI analyses, including segmentation, especially if these are to be automated or use atlases. Automated registration techniques have become routine in many areas, especially in neuroscience, because of the relative ease of registering the head and its constituents afforded by their comparative rigidity and the presence of numerous morphologic features. The three main uses of MR image registration are registration of data from multiple modalities (e.g., PET–MRI) in a given subject to aid interpretation and quantification; registration of scans from individual subjects acquired at different times to highlight change (e.g., longitudinal imaging studies) and eliminate undesired artifacts (e.g., motion correction in fMRI); and registration of data from different subjects. In epilepsy studies, segmented MR images can be registered to PET images to correct for partial volume effects due to the presence of CSF 8. The registration of fMRI and structural MRI data also is essential in the interpretation of the functional data (Fig. 3). Serial MR images can be registered precisely, markedly improving volumetric reproducibility and sensitivity to change 9(Fig. 2). Finally, automatic region-of-interest (ROI) definition based on digital atlas matching through the registration of a representative brain to individual subject's images is emerging as a practical solution to localization problems 10. Illustration of multimodality image registration. The head is shown from a T1-weighted volume scan. A BOLD activation derived from a spike-triggered echo-planar functional magnetic resonance imaging acquisition is shown in the "hot metal" colors. The blue arrows represent dipoles fitted to 64-channel EEG recorded on the same patient. Voxel and volume-based MRI data analyses allow access to otherwise elusive, biologically important information about refractory epilepsy. The poor outcome from resective surgery for malformations of cortical development (MCD) has been attributed to widespread unresected pathology, but pathologic proof is rare. MCD usually disrupts cortical organisation, a cerebral attribute of unexpected invariance. Methods detecting deviation from normal cortical volumetric and voxel-based parameters might thus reveal occult structural abnormalities, providing information for surgical management, prognostication, and for aetiologic classification. Volume measures may be directed at specific regions, exemplified by hippocampal volumetry, or at the brain as a whole. Regional measures are prone to numerous biases, especially when the ROI is difficult to define, whereas global measures are complicated by issues of intersubject homology and registration, and may be relatively insensitive. However, MRI volumetry promises access to previously unobtainable biologic information, and so merits development despite its limitations. The "block" technique was designed to identify such abnormalities using ROI methods. Within arbitrary but reproducible coronal blocks of GM or WM, each extending a fixed proportion of the anteroposterior extent of the hemisphere, the proportional regional distribution of GM and WM is measured 11. An initial study examined subjects treated surgically for hippocampal sclerosis (HS). Results suggested that extrahippocampal quantitative changes correlated with poor outcome after temporal lobectomy, suggesting that the method identified changes of biologic significance 12. However, the method was time consuming, laborious, and semiautomated. The method is now fully automated, and is being used in a prospective study of 100 consecutive HS cases. Correlative studies with postmortem pathology also have been undertaken: in pilot studies, block abnormality findings generated a neuropathologic sampling strategy that revealed pathology not identified from gross brain examination in a subject with tuberous sclerosis. Further studies are required. The block method is robust and able to analyze grossly abnormal brains. It groups together large voxel numbers, however, and may thus not detect smaller regions of abnormality. Voxel-based morphometry (VBM), designed to study PET data, has been applied to structural MRI. Using automated normalization, segmentation, and statistical estimation, groups of subjects may be compared to determine whether a given averaged voxel contains unexpected signal, that is, whether a given voxel is more or less likely to be GM or WM. VBM has been used in many quantitative studies in epilepsy, for example, showing changes in the frontal lobes of some patients with juvenile myoclonic epilepsy 13, suggesting a possible underlying structural basis for epilepsy in these cases, in keeping with some pathologic data 14, and also suggesting that the electroclinical syndrome may be heterogeneous, as supported by neurogenetic studies 15. However, the limitations of VBM must be borne in mind. In particular, the statistical bases require careful consideration if oversimplifications and errors are to be avoided. A recent review of the application of VBM to the nonrandom (stationary) data from structural MRI is essential reading for those contemplating using VBM 16. Volumetry also may be applied to specific ROIs. The outstanding example is the hippocampus. Other areas of the brain may also merit mensuration. The human mesial temporal region consists of hippocampus, amygdala, and the parahippocampal region; the latter is itself subdivided into entorhinal cortex (EC), perirhinal cortex (PC) and posterior parahippocampal cortex (PPC; areas TH and TF) 17. In early studies of temporal lobe resection specimens, the term "mesial temporal sclerosis" was introduced to describe widespread pathologic changes encompassing hippocampus, amygdala, and the parahippocampal region 18. More recently, MRI in temporal lobe epilepsy (TLE) has concentrated on the hippocampus. Given pathologic observations, however, in vivo MRI volume changes of different parahippocampal subfields might be of biologic significance. A recent study reported volumetric measurements of these structures in 25 patients with intractable TLE and unilateral hippocampal atrophy compared with normal controls. In this study, MRI volumetric images were automatically registered into stereotaxic space 19 to adjust for differences in total brain volume and brain orientation and to facilitate the identification of boundaries by minimizing variability in slice orientation 20. In addition, each image underwent automated correction for intensity nonuniformity due to radiofrequency inhomogeneity, with intensity standardization 21. The hippocampus, EC, PC, and PPC were segmented manually using mouse-driven software according to previously described protocols 22-25. Interestingly, both the ipsilateral EC and PC were smaller in patients than in normal controls (p < 0.001). Individual analysis showed that the majority of patients had abnormal EC, 72% of whom had atrophy ipsilateral to the seizure focus. Only five of 25 patients had ipsilateral PC atrophy. Thus, in patients with intractable TLE and unilateral hippocampal atrophy, there is decreased volume of the parahippocampal region ipsilateral to the seizure focus. However, this atrophy is unevenly distributed. The EC was almost always abnormal, the PC was sometimes abnormal, and the PPC was always normal. In vitro studies of focal epileptogenesis in combined hippocampal–entorhinal slices show that the EC possesses an intrinsic capacity to generate epileptiform discharges 26. After amino-oxyacetic acid injection in the rat EC, there is extensive cell loss in layer III 27 of medial EC identical to changes found in human TLE 28. EC damage may contribute to long-lasting changes in excitability in the EC and the hippocampus, and play a primary role in genesis and spread of temporal lobe seizures 29. The reason for preferential damage to the EC in these patients must be further explored. This information might eventually be used in more sophisticated surgical planning. The approach here illustrates the potential for specific ROI studies to expand our understanding of the substrate of refractory epilepsy. Morphology and texture are important features for visual image assessment. Computer-based texture analysis of digital images provides quantitative information about spatial gray-level variations in pixel neighborhoods 30-32, On MRI, focal cortical dysplasia (FCD) is characterized by variable cortical thickening, a poorly defined GM–WM transition, and hyperintense signal within the dysplastic lesion with respect to normal cortex 33. Whereas high-resolution MRI permits identification of FCD in many patients 2, 34, many FCD cases are characterized by minor structural abnormalities that are too subtle to be detected by inspection, reformatting, or volumetry. Given that the brain has many properties, and different combinations of these properties are thrown into relief by different methods, it is not surprising that some lesions may be detected only by using methods directed against specific lesion characteristics. Thus voxel-based image-processing techniques focusing on pixel intensities, local intensity gradients, and GM thickness may specifically identify focal developmental lesions. In a recent study, patients who had histologically proven FCD after surgery were studied using methods highlighting these features. Preoperative volumetric images were acquired using a T1-fast field echo sequence. Images were analyzed using software developed in the Montreal Neurological Institute. Segmentation (GM or WM) was undertaken using a histogram-based method with automated threshold. Image-processing features were calculated for each individual voxel within the T1-weighted 3D MRI, resulting in a 3D map for each feature. To model cortical thickening, a morphologic operator was used wherein each individual voxel was used as the starting point for GM extent (run-length coding), measured in each possible point-to-point direction. To model GM–WM transition blurring, the absolute gradient of gray-level intensities, a first-order texture feature, was calculated. To model the hyperintense signal within the dysplastic lesions, a feature that calculates the absolute difference between the intensity of a given voxel and the intensity at the GM–WM boundary was devised. To maximize visibility of FCD lesions, a ratio map (GM thickness × relative intensity/gray-level intensity gradient) was generated. A series of images consisting of MRIs and ratio maps for 16 patients and 20 healthy control subjects was presented in random order to two trained observers who were unaware of the final diagnosis. Overall accuracy (correctly classified/total cases) was 91.7% for the ratio maps and 77.8% for the raw MRI. Sensitivity was 87.5% for the ratio maps compared with 50% for MRI (p < 0.003, Pearson's χ2). Specificity was 95% for ratio maps and 100% for MRI. Cohen's κ was 0.53 for MRI, indicating moderate agreement, and 0.83 for ratio maps, indicating strong agreement beyond chance between the two observers. This study, using voxel-based image postprocessing methods modeled on known MRI features of FCD, revealed the possibility of increasing sensitivity of lesion detection by 37.5% over conventional MRI analysis while accuracy was increased by 15%. Although some subtle cortical lesions are being increasingly recognized using routine MRI and multiplanar 2, 35 and curvilinear 5 reformatting, these results indicate that detection of subtle dysplastic lesions may be further improved by performing quantitative analysis of the structural changes that characterize FCD pathologically and in vivo on MR images. After the detection of lesions, using routine or advance MRI methods, to allow precise surgical localization, it is necessary to map out regions of normal and abnormal brain structure and activity and carry that roadmap into the surgical operating environment, while constantly checking for positional and functional accuracy. The information can then be used to guide resection of the epileptogenic focus. In recent years, there has been a move toward using preoperatively acquired structural and functional image information as a visualization guide for this task, with the preoperative information registered to the intraoperative environment, typically using sets of markers. However, this approach has a variety of problems that must be solved in the next few years to keep image-guided surgery technology viable. The problems include a number of systems engineering issues, including fully fusing all of the available pre- and intraoperative functional information into the intraoperative displays, accounting for brain shift after opening the dura, and providing a fully interactive visualization environment that allows the accurate delineation and labeling of neuroanatomic structure and function. Some recent efforts have concentrated on three specific areas related to these problems: (a) the segmentation and measurement of cortical GM, (b) compensation for brain shift during image-guided neurosurgery, and (c) localization, visualization, and measurement of implanted electrode arrays. Segmentation can be undertaken in many ways (see also earlier); it provides basic constraints for the brain-shift algorithm described later. One approach uses an automated coupled-level set strategy 36. In this approach, several seed points are placed anywhere in the WM using midbrain image slices. Then the algorithm runs to convergence, simultaneously solving two coupled differential equations that solve for the gray/white and gray/CSF boundaries in three dimensions in the volumetric high-resolution anatomic MRI dataset. In an effort to account for brain-shift error in the current implementation of image-guided neurosurgery systems, a biomechanical model–based approach to brain-shift compensation has been developed 37. Experiments have shown that brain shift can be upwards of 5–7 mm, even with opening of the dura alone (due only to loss of fluid and gravitational effects). In addition the physical changes in the brain caused by electrode implantation and removal before the definitive surgery makes brain shift a dominant source of positional error between presurgically acquired image information and that displayed in the intraoperative navigation system. One approach to brain-shift compensation is to deform a preoperatively constructed model by using intraoperatively acquired positional data from video cameras and then to display the deformed versions of the preoperative images for navigation. The modeling equations are solved with finite-element methods. In the last few years, it has become increasingly useful to integrate the position of intracranial recording electrodes with functional and structural data. However, electrode artifact has been a limiting factor. It is now possible to correct for the effects of MR signal artifact generated by implanted electrodes by extracting the 3D evidence of the approximately spherical artifacts created by arrays of surgically implanted electrode grids used to monitor epileptic activity. One approach uses a simple physical model of a deformable, but nonstretchable, Mylar sheet in which the electrodes are embedded. This model then constrains the solution and helps reliably recover electrode positions. The physical modeling is performed by including distance-preserving constraints with a nonlinear spring network in which the net is always relaxed so that the springs reach their rest lengths. By reconstructing these grids, and then mapping the information back to the preoperative state with brain-shift compensation strategies, it is possible to display the electrode information in the context of a variety of preoperative imaging data 38. This should permit the accurate study of EEG data from the intraoperatively placed electrodes in relation to a variety of preoperative functional and structural information. Many methods have been devised to increase the yield of information from MRI data. The ultimate tests of the utility of these methods are surgical cure, pathologic confirmation of epileptogenic pathology, or correlation with genetic etiology and animal models. The brain is extremely complex, and no one method of analysis is likely to be able to describe the brain or identify abnormalities within its structure adequately, as different sequences are commonly used in the study of different diseases: the method used to study it should be suited to the nature of the changes being sought. In some cases, concordance of results from multiple analytic methods will occur, and this must serve to increase confidence in the findings, at least until such time as these methods become gold standards in themselves rather than in reference to genetic, pathologic, or surgical outcome data. Methods devised need comparison preoperatively in the same patients, and free and public access to all the methods for the entire epilepsy surgery community is an important part of this process, similar to the deposition of gene sequences in databanks with acceptance for publication of each new genetic discovery.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.